tulokset
Silmäile
Osuva
Osuva on Vaasan yliopiston avoin julkaisuarkisto. Osuva sisältää Vaasan yliopiston omat julkaisut, opinnäytteet ja tieteellisten artikkeleiden rinnakkaistallenteet. Osuvaan sisältyy julkaisujen viitetietoja, tiivistelmiä ja kokotekstejä. Sähköisten arkistokokoelmien sisältö ei ole luettavissa verkossa.
Kokoelmat Osuvassa
Viimeksi tallennetut
Trustworthy LLMs for Ethically Aligned AI-based Systems: A PhD Research Plan
de Cerqueira, José Antonio Siqueira; Rousi, Rebekah; Xi, Nannan; Hamari, Juho; Kemell, Kai Kristian; Abrahamsson, Pekka; M., Deekshitha (toim.); Santos, Rodrigo (toim.); Khanna, Dron (toim.); Elshan, Edona (toim.) (RWTH Aachen, 2025)
In response to growing concerns around trustworthiness and ethical alignment in AI systems, this PhD aims to investigate how Large Language Models (LLMs) can be leveraged to support ethically aligned AI development in software engineering. Despite advancements, integrating ethical principles into AI workflows remains challenging, particularly in real-world applications that require compliance with emerging regulations, such as the EU AI Act. We will develop a Visual Studio Code (VSCode) Generative AI (GenAI) Extension powered by a multi-agent LLM system with Retrieval-Augmented Generation (RAG) capabilities. The extension will be designed to aid developers by evaluating code compliance with ethical standards, providing actionable recommendations to embed trustworthiness from early stages of development. The GenAI Extension will be evaluated through an iterative design science approach, encompassing dataset generation, ethical benchmarking, and practitioner testing. A dataset of over 2000 ethically aligned AI systems, will be created in compliance with leading regulatory frameworks, serving as a foundation for this tool’s assessments. With this work, we hope to assist developers, particularly in startups and SMEs, by providing practical resources for building ethically aligned AI within limited resources. Through this approach, we aim to bridge the gap between abstract ethical principles and actionable software development practices, making ethical AI more accessible across industry contexts.
Grounded Ethical AI: A Demonstrative Approach with RAG-Enhanced Agents
de Cerqueira, José Antonio Siqueira; Khan, Ayman Asad; Rousi, Rebekah; Xi, Nannan; Hamari, Juho; Kemell, Kai-Kristian; Abrahamsson, Pekka; M., Deekshitha (toim.); Santos, Rodrigo (toim.); Khanna, Dron (toim.); Elshan, Edona (toim.) (RWTH Aachen, 2025)
Large Language Models (LLMs) have become central in various fields, yet their trustworthiness remains a pressing concern, especially in developing ethically aligned AI-based systems. This paper presents a demonstration of an LLM-based multi-agent system incorporating Retrieval-Augmented Generation (RAG) to support developers in creating AI systems that align with legal and ethical guidelines. Leveraging documents like the EU AI Act, AI HLEG guidelines, and ISO/IEC 42001:2024, the prototype utilizes multiple agents with specialized roles, structured conversations, and debate rounds to enhance both ethical rigor and trustworthiness. Initial evaluations on real-world AI incidents reveal that this system can produce AI solutions adhering to specific ethical requirements, though further refinements are needed for citation accuracy and practical application. This demonstration illustrates the potential of RAG-enhanced LLMs to operationalize AI ethics and regulatory compliance within the development process, highlighting future directions for achieving more reliable and ethically robust AI solutions.
Boreal Forest Fire: UAV-collected Wildfire Detection and Smoke Segmentation Dataset
Pesonen, Julius; Raita-Hakola, Anna-Maria; Joutsalainen, Jukka; Hakala, Teemu; Akhtar, Waleed; Koivumaeki, Niko; Markelin, Lauri; Suomalainen, Juha; Alves de Oliveira, Raquel; Polonen, Ilkka; Honkavaara, Eija (Springer, 2025)
Automated image-based wildfire detection suffers from a lack of open-access data, especially data with annotations. Our dataset targets the gap by providing human and computer vision foundation-model co-annotated images from an uncrewed aerial vehicle (UAV) perspective from Finnish boreal forest environments. The images and videos were collected at multiple prescribed burning events, and the data were used to successfully train wildfire detection models in our previous studies, proving their value for the task. The Boreal Forest Fire dataset contains three sections: images with bounding box annotations, video clips with labels and images with segmentation masks. Alongside the data, we have released code, ensuring that the data is simple to use.
From Healthcare Technology to Care Robot-Literate Practitioners
Turja, Tuuli; Kork, Anna-Aurora; Andrikopoulou, Elisavet (toim.); Gallos, Parisis (toim.); Arvanitis, Theodoros N. (toim.); Austin, Rosalynn (toim.); Benis, Arriel (toim.); Cornet, Ronald (toim.); Chatzistergos, Panagiotis (toim.); Dejaco, Alexander (toim.); Dusseljee-Peute, Linda (toim.); Mohasseb, Alaa (toim.); Natsiavas, Pantelis (toim.); Nakkas, Haythem (toim.); Scott, Philip (toim.) (IOS Press, 2025)
Current forms of health technology literacies fail to fully address the multifaceted nature of care robot literacy (CRL). As an occupational asset for healthcare practitioners, CRL involves the ability to use and interact with mobile, artificially intelligent (AI) -driven mechatronic devices within the working environment. A variety of new generation technologies are introduced in healthcare. However, occupational and ethical standards need to be in line with utilization of any novel technologies. Based on a synthesis on existing literature, this poster will discuss the socio-temporal preconditions and further steps to develop CRL among healthcare practitioners.
Scalable Consensus Algorithm and Storage in Decentralized Blockchain (SCSB) For Heterogeneous Internet of Things (IoT) Systems
Singh, Inderpal; Singh, Balraj; Faheem, Muhammad (Institution of engineering and technology, 2025)
The internet of things (IoT) is widespread in various real-time applications in developing smart environments. The involvement of numerous network devices and users in the system includes illegitimate and malicious entities. Also, the participation of numerous devices leads to scalability issues. The decentralized blockchain is one of the promising solutions to satisfy all the requirements of security. In this paper, a priority-based lightweight authentication and access control is designed for cloud-IoT (SCSB) with the assistance of decentralized blockchain technology. The SCSB design consists of the data owner (DO), data user (DU), trusted authority (TA), and cloud server. The cloud server manages a huge amount of data, hence, it receives multiple DU requests. The DO is also authenticated and then allowed to upload data. The data in the cloud is clustered and then stored, which is scalable in storage. In this work, density-based spatial clustering applications in noise (H-DBSCAN) is presented with a hybrid distance measure. The validation of multiple requests into the blockchain is conducted using novel proof-of-authentication, which selects a trusted node for validation. To ensure secure data upload, the priority, i.e., the confidentiality level of data, is predicted from the dual fuzzy algorithm, and then the data is secured. For a high confidential level and low confidential level, a lightweight TWINE algorithm and differential privacy are incorporated. The use of lightweight algorithms and parallel processing algorithms in dual fuzzy enables it to operate with numerous devices while utilizing limited resources and time, which solves the scalability issue. The proposed SCSB shows better performance results than the previous research algorithms.